What I’ve Learned From Agentic Design • Chris LaChance

Skip to main content

After 6 months of investing heavily into AI-driven design (e.g. Agentic Design), I thought this article would serve as a help for others and benchmark for myself for end of year reflection.

What It Looks Like Right Now

My previous workflow was coordinating with both Project Managers (PM), Stakeholders, Users, other Designers and Devs; taking those requirements into the medium that presented the solution best for sharing execution; and refining with teams until live. Typically that was a combination of Figma for static documented assets and / or an HTML prototype or exploration shared out via Codepen, but sometimes that was also just tweaking a proposal from a PM.

Honestly, not much has changed in this new world except the tooling. Things still require documentation and sometimes live prototypes. Have had some great success at connecting Figma MCP to Copilot running a Claude model in order to generate accessible 3:1 color series for our data viz patterns. I use TypeWhisper (a local LLM that converts voice to text) A LOT. So most of the time I’m talking to AI instead of typing (since we’ve come full circle and are now essentially back in terminal doing design…) End of the day, its still the right tool at the right time for the right job.

AI is rocking it for hammering out & testing new complex interactions, but typically can’t nail that last 20% of polish that’s critical in UI and interaction design without a serious amount of prompting (or manual intervention). It’s also bad at documentation (or my process needs to change is more likely).

Figma Make is WAY too expensive, and our team has jettisoned it in place of Copilot, where I paired with a dev to set up repo where vibed prototypes can be reviewed only on our intranet, completely replacing the Figma Make process (except commenting, for now).

The parts of the job that are a drain, such as managing email, schedules, incoming messages, etc. still need some serious AI optimization.

What I’m Seeing Generally

I don’t want to be either naive, nor pessimistic here. There are some serious “Geez, that’s amazing” moments, and just as many “What are you doing!?” moments as well.

Pros

Giving power to anyone to tell their story via an interactive design is really just amazing. I’ve always taken the Ratatouille approach to design “Anyone can cook”. Awesome solutions come from anywhere – PMs, Devs, Interns, Users. I find this invigorating, not intimidating. Seeing designers and awesome PMs bring great ideas to life greatly improves the product.

Automation has so much potential and I’m looking forward to more of it, especially for the boring stuff. Tools like TypeWhisper, text expansion, automated summaries are saving me colossal amounts of time, while keeping my responses even more personal. It’s really great.

Being able to auto-code up math, design standards, and accessibility principles into vibed Figmas and demos is really just spectacular.

There are so many opportunities for localization, accessibility, and other incredibly empowering tools for ourselves, partners, and clients – it’s really quite amazing.

Cons

Agentic tooling is exponentially stoking the fire of cutting corners, offloading thinking, and bad habits. Every prompt box tells us not to trust it – but its seems so many trust it more than even their Subject Matter Experts (SMEs) and users these days. I recently trusted that the documentation generated by the LLM in a Figma was accurate, but when a dev went to implement, the colors on the boxes didn’t match the hex codes typed out in the docs. I missed it, trusting the AI way more than I should have.

The empowerment of AI seems to turning people to their basest instincts of “Give me what I want, NOW” without regarding consequences.I’ve likened this effect to toddler jamming a metal fork in to an electrical socket because “It seems right to me!” Telling them “No”, just empowers them to vibe it up anyways when you aren’t looking, so the job has become a bit of ‘switching the metal fork for a plastic one’ before they do… Smart defaults, yet allowing “Not recommended” options is one way I’m considering resolving this.

The power also means that people seem to feel less & less of a need to coordinate with anyone else before shipping (or its become secondary) – creating colliding, inconsistent, and surprising experiences. And customers, clients, and partners don’t really like surprises. More, now than ever – design work must be ‘align work’.

AI / LLM vibe work reduces how much brainpower & skill is required to get powerful results – a double-edged sword. That striving is part of what makes work rewarding. In addition, flow state now becomes nigh impossible to reach while babysitting an LLM every 5 minutes for review, waiting, ‘approve this’, and thinking of next steps. I am trying to fill that time with responding to messages, filing bugs, upskilling, reading web news,...